Autonomous robotic systems, like autonomous vehicles and robotic search and rescue, require efficient on-device training for continuous adaptation of Deep Reinforcement Learning (DRL) models in dynamic environments. This research is fundamentally motivated by the need to understand and address the challenges of on-device real-time DRL, which involves balancing timing and algorithm performance under memory constraints, as exposed through our extensive empirical studies. This intricate balance requires co-optimizing two pivotal parameters of DRL training -- batch size and replay buffer size. Configuring these parameters significantly affects timing and algorithm performance, while both (unfortunately) require substantial memory allocation to achieve near-optimal performance. This paper presents R^3, a holistic solution for managing timing, memory, and algorithm performance in on-device real-time DRL training. R^3 employs (i) a deadline-driven feedback loop with dynamic batch sizing for optimizing timing, (ii) efficient memory management to reduce memory footprint and allow larger replay buffer sizes, and (iii) a runtime coordinator guided by heuristic analysis and a runtime profiler for dynamically adjusting memory resource reservations. These components collaboratively tackle the trade-offs in on-device DRL training, improving timing and algorithm performance while minimizing the risk of out-of-memory (OOM) errors. We implemented and evaluated R^3 extensively across various DRL frameworks and benchmarks on three hardware platforms commonly adopted by autonomous robotic systems. Additionally, we integrate R^3 with a popular realistic autonomous car simulator to demonstrate its real-world applicability. Evaluation results show that R^3 achieves efficacy across diverse platforms, ensuring consistent latency performance and timing predictability with minimal overhead.

The growing focus on indoor robot navigation utilizing wireless signals has stemmed from the capability of these signals to capture high-resolution angular and temporal measurements. Prior heuristic-based methods, based on radio frequency propagation, are intuitive and generalizable across simple scenarios, yet fail to navigate in complex environments. On the other hand, end-to-end (e2e) deep reinforcement learning (RL), powered by advanced computing machinery, can explore the entire state space, delivering surprising performance when facing complex wireless environments. However, the price to pay is the astronomical amount of training samples, and the resulting policy, without fine-tuning (zero-shot), is unable to navigate efficiently in new scenarios unseen in the training phase. To equip the navigation agent with sample-efficient learning and {zero-shot} generalization, this work proposes a novel physics-informed RL (PIRL) where a distance-to-target-based cost (standard in e2e) is augmented with physics-informed reward shaping. The key intuition is that wireless environments vary, but physics laws persist. After learning to utilize the physics information, the agent can transfer this knowledge across different tasks and navigate in an unknown environment without fine-tuning. The proposed PIRL is evaluated using a wireless digital twin (WDT) built upon simulations of a large class of indoor environments from the AI Habitat dataset augmented with electromagnetic (EM) radiation simulation for wireless signals. It is shown that the PIRL significantly outperforms both e2e RL and heuristic-based solutions in terms of generalization and performance. Source code is available at \url{https://github.com/Panshark/PIRL-WIN}.

Pre-training a diverse set of neural network controllers in simulation has enabled robots to adapt online to damage in robot locomotion tasks. However, finding diverse, high-performing controllers requires expensive network training and extensive tuning of a large number of hyperparameters. On the other hand, Covariance Matrix Adaptation MAP-Annealing (CMA-MAE), an evolution strategies (ES)-based quality diversity algorithm, does not have these limitations and has achieved state-of-the-art performance on standard QD benchmarks. However, CMA-MAE cannot scale to modern neural network controllers due to its quadratic complexity. We leverage efficient approximation methods in ES to propose three new CMA-MAE variants that scale to high dimensions. Our experiments show that the variants outperform ES-based baselines in benchmark robotic locomotion tasks, while being comparable with or exceeding state-of-the-art deep reinforcement learning-based quality diversity algorithms.

We propose a new algorithm for variance reduction when estimating $f(X_T)$ where $X$ is the solution to some stochastic differential equation and $f$ is a test function. The new estimator is $(f(X^1_T) + f(X^2_T))/2$, where $X^1$ and $X^2$ have same marginal law as $X$ but are pathwise correlated so that to reduce the variance. The optimal correlation function $\rho$ is approximated by a deep neural network and is calibrated along the trajectories of $(X^1, X^2)$ by policy gradient and reinforcement learning techniques. Finding an optimal coupling given marginal laws has links with maximum optimal transport.

The world's leading Gran Turismo 7 players have had their shot at beating Sony's AI racer, GT Sophy, and they lost spectacularly. Now, the bot has added drifting to its move set to rub salt on the wounds, as seen in a video from the Gran Turismo World Series 2023 event. The video clearly shows an AI-controlled vehicle drifting around the track like an absolute maniac. Meanwhile, most people can't pull off successful drifts in Mario Kart, let alone in realistic racers like Gran Turismo. GT Sophy, you can guess what the GT stands for, is the result of more than six years of development between Sony AI and Sony Interactive Entertainment, and utilizes deep reinforcement learning methods.

This work uses the entropy-regularised relaxed stochastic control perspective as a principled framework for designing reinforcement learning (RL) algorithms. Herein agent interacts with the environment by generating noisy controls distributed according to the optimal relaxed policy. The noisy policies on the one hand, explore the space and hence facilitate learning but, on the other hand, introduce bias by assigning a positive probability to non-optimal actions. This exploration-exploitation trade-off is determined by the strength of entropy regularisation. We study algorithms resulting from two entropy regularisation formulations: the exploratory control approach, where entropy is added to the cost objective, and the proximal policy update approach, where entropy penalises policy divergence between consecutive episodes. We focus on the finite horizon continuous-time linear-quadratic (LQ) RL problem, where a linear dynamics with unknown drift coefficients is controlled subject to quadratic costs. In this setting, both algorithms yield a Gaussian relaxed policy. We quantify the precise difference between the value functions of a Gaussian policy and its noisy evaluation and show that the execution noise must be independent across time. By tuning the frequency of sampling from relaxed policies and the parameter governing the strength of entropy regularisation, we prove that the regret, for both learning algorithms, is of the order $\mathcal{O}(\sqrt{N}) $ (up to a logarithmic factor) over $N$ episodes, matching the best known result from the literature.

Reinforcement learning algorithms typically struggle in the absence of a dense, well-shaped reward function. Intrinsically motivated exploration methods address this limitation by rewarding agents for visiting novel states or transitions, but these methods offer limited benefits in large environments where most discovered novelty is irrelevant for downstream tasks. We describe a method that uses background knowledge from text corpora to shape exploration. This method, called ELLM (Exploring with LLMs) rewards an agent for achieving goals suggested by a language model prompted with a description of the agent's current state. By leveraging large-scale language model pretraining, ELLM guides agents toward human-meaningful and plausibly useful behaviors without requiring a human in the loop. We evaluate ELLM in the Crafter game environment and the Housekeep robotic simulator, showing that ELLM-trained agents have better coverage of common-sense behaviors during pretraining and usually match or improve performance on a range of downstream tasks. Code available at https://github.com/yuqingd/ellm.

We propose a physics-based method for synthesizing dexterous hand-object interactions in a full-body setting. While recent advancements have addressed specific facets of human-object interactions, a comprehensive physics-based approach remains a challenge. Existing methods often focus on isolated segments of the interaction process and rely on data-driven techniques that may result in artifacts. In contrast, our proposed method embraces reinforcement learning (RL) and physics simulation to mitigate the limitations of data-driven approaches. Through a hierarchical framework, we first learn skill priors for both body and hand movements in a decoupled setting. The generic skill priors learn to decode a latent skill embedding into the motion of the underlying part. A high-level policy then controls hand-object interactions in these pretrained latent spaces, guided by task objectives of grasping and 3D target trajectory following. It is trained using a novel reward function that combines an adversarial style term with a task reward, encouraging natural motions while fulfilling the task incentives. Our method successfully accomplishes the complete interaction task, from approaching an object to grasping and subsequent manipulation. We compare our approach against kinematics-based baselines and show that it leads to more physically plausible motions.

We present a novel approach to address the challenge of generalization in offline reinforcement learning (RL), where the agent learns from a fixed dataset without any additional interaction with the environment. Specifically, we aim to improve the agent's ability to generalize to out-of-distribution goals. To achieve this, we propose to learn a dynamics model and check if it is equivariant with respect to a fixed type of transformation, namely translations in the state space. We then use an entropy regularizer to increase the equivariant set and augment the dataset with the resulting transformed samples. Finally, we learn a new policy offline based on the augmented dataset, with an off-the-shelf offline RL algorithm. Our experimental results demonstrate that our approach can greatly improve the test performance of the policy on the considered environments.

This paper aims at the algorithmic/theoretical core of reinforcement learning (RL) by introducing the novel class of proximal Bellman mappings. These mappings are defined in reproducing kernel Hilbert spaces (RKHSs), to benefit from the rich approximation properties and inner product of RKHSs, they are shown to belong to the powerful Hilbertian family of (firmly) nonexpansive mappings, regardless of the values of their discount factors, and possess ample degrees of design freedom to even reproduce attributes of the classical Bellman mappings and to pave the way for novel RL designs. An approximate policy-iteration scheme is built on the proposed class of mappings to solve the problem of selecting online, at every time instance, the "optimal" exponent $p$ in a $p$-norm loss to combat outliers in linear adaptive filtering, without training data and any knowledge on the statistical properties of the outliers. Numerical tests on synthetic data showcase the superior performance of the proposed framework over several non-RL and kernel-based RL schemes.